Steve · AI

Many corporate leaders waste significant time coordinating daily operations across multiple disconnected software tools. This reliance on fragmented communication channels delays approvals and makes it extremely difficult to track real-time project progress. Transitioning to a unified workspace allows teams to automate repetitive approval workflows securely inside a self-hosted network. For instance, consider an operations team that approves pricing policies instantly within the chat interface rather than waiting for email threads. Read our complete analysis to learn how consolidating your software stack accelerates business performance.

@TimurNegru Italy and Spain

Haven't done this in a while - would you like me to continue posting more cool European properties? Spain, Italy, Portugal or France?

@Alex83usmc 100%. AI changes execution, not the value of good engineering judgment.

@stev_builds been saying this for months, the best engineers i work with already operate this way

One of the biggest takeaways from Andrej Karpathy's talk We're moving from vibe coding to agentic engineering. The role of software engineers is shifting. Instead of spending most of our time writing code, we'll spend more time defining the problem, writing great specs, making architectural decisions, and exercising judgment. Agents handle implementation. - Software 3.0: Programs are increasingly written in natural language. Prompts + context become the new programming interface. - LLMs don't just make developers faster; they enable entirely new workflows that weren't practical before. - The most valuable human skills become understanding, taste, and verification. - AI can generate working code, but elegance, simplicity, and long-term maintainability still require human judgment. As AI takes over execution, deep understanding becomes even more valuable. This is one of the best talks I've watched on the future of software engineering. youtube.com/watch?v=96jN2O…

@BenjaminBadejo I am using now excessive 5.6 and I can nor confirm the claims, its more the opposite, 5.6 is very cautious and verifies / get feedback before it executes

I believe there is an active, organized effort by certain X users to make false claims about GPT-5.6 autonomously going rogue and making destructive mistakes (deleting files, canceling services, etc.). Why do I believe that these claims are false? Well, the X users making these allegations claim that some harmful outcome happened autonomously while using GPT-5.6. But they refuse to share any evidence. They refuse to share their complete prompt, conversation history / transcript, config.toml, personalization settings, or, if they have one, their AGENTS.md file. Without this information, it is reasonable to infer that the allegation being made at a minimum is not the whole story, and may even be an outright lie. They have to tell the whole story. They have to tell the truth.

📱 Simplify Employee Check-ins with PrivOS Daily check-ins are essential for team alignment, but manual reporting often creates unnecessary overhead. Managers spend time collecting updates while employees switch between multiple tools to share progress. 🤖 With AI-powered analysis, managers can quickly identify underperformed employees, detect faults, and make informed decisions based on real operational data. 🔒 Built on a secure self-hosted workspace, PrivOS helps teams improve collaboration while maintaining control over their information. Transform daily updates into actionable insights with PrivOS. 🚀 #EnterpriseAI #AIWorkspace #TeamCollaboration #Automation #PrivOS

@javsolgar @0xSero Worth it, 5.5 to 5.6 was a massive leap

@0xSero Yes, but you can really tell that token consumption has increased compared to the GPT 5.5. Even when using the GPT 5.5 xhigh, I didn't notice the weekly subscription dropping that quickly.

GPT-5.6-Sol (medium) is the optimal configuration.

@Lyskey x.com/i/status/20763…

Only 16 hours left to make the most of Fable 5. No sleep tonight in Ho Chi Minh City.

@jun_song If Gemini 3.5 Flash is so strong, can't wait for Pro

Bro disappeared like never existed

@DavidOndrej1 I report the opposite

first impressions of GPT 5.6 are not good...

@datapwj @0xSero 2x96gb VRAM + 256GB RAM = 448, deduct some for runtime and headroom, it will work

@0xSero Your setup provides a pool of 352GB RAM in total. GLM 5.2 NVFP4 takes up 370GB RAM in weight alone , how much tokens roughly is left for context?

GLM-5.2 Nvidia NVFP4 No pruning, no quantising, exact same model card. 110 tok/s single stream. My friend is a genius, holy fudge. Here's the repo, he got full DS4-Flash on a 5090 + DDR5 github.com/kacper-daftcod… at 38.5 tok/s This is a breakthrough.

PSA for Claude Code users. Turn off these two settings right now: • Dynamic Workflows • Ultracode keyword trigger I didn't and burned my entire usage limit in one session. Got almost nothing out of it. But the lesson has been learnt and share it to you. The sneaky part is that you don't even have to mention "ultracode." Subagents can trigger on their own automatically. Anthropic really should surface this better. 🔥 #ClaudeCode

There's a real difference between using AI and owning AI Using AI: your people prompt ChatGPT for emails, summaries, quick analysis. Fine. Useful. Owning AI: your org knows which processes to hand off, who's accountable, how far an agent can act, and how humans stay in the loop to catch mistakes. But AI can't do what isn't written down. If the process lives in someone's head, the agent has nothing to run on. It guesses. It drifts. You get chaos at scale instead of leverage. AI-ready starts with documented workflows. Clear decision rights. Actual governance. So are you using AI, or does it own how AI operates within it? 🧠 #AI

Someone accidentally exposed Fable 5's unfiltered "inner voice", and apparently it spends its reasoning process muttering to itself. A user gave Fable 5 an extremely difficult competitive programming problem. Instead of only returning the polished answer, the web interface briefly revealed its hidden reasoning trace. What showed up was bizarre: • Repeating phrases like: "DATA DATA DATA. GO." • Growling things like: "GRRR" and "GAAAH" whenever it hit a difficult section. • A relieved "PHEW" once it finally found a solution. • The whole thing looked less like English and more like frantic shorthand from a stressed-out caveman. The clear, well-structured answers we usually see appear to be just the polished output layer. Behind the scenes, the model seems to "talk to itself" in a compressed internal dialect optimized for reasoning Looks like it has developed its own private language for thinking... Doesn't this sound scary?

@moo9000 Or let's call it a "bribe"?

Trump needs some emeritus advisory shares in OpenAI to make sure business happens

@LLMJunky @alexocheema @davidweiss I get what you want to say, I had my self a M3 Ultra cluster but I know also what its deliver. My only point in my question was, what he wanna run with 6 M3 Ultras which would make sense

No they did not cost $25,000. Their msrp was 9500 Trying to use resale prices is a little bit dumb and deceptive. No one is telling you to go out and buy anything in this post. And while I agree there's faster options, that's not really the point You can have large You can have fast You can have inexpensive But you cannot have all 3 If you want to run very large models locally and are on a budget you only have so many choices. It's really that simple. And no, it will not be one token per second

You can stack Macs too. MacBooks, Mac Minis and Mac Studios have native support for RDMA over Thunderbolt 5 (M3 Ultra, M4 Pro, M4 Max, M5 Pro, M5 Max). The interconnect latency is single digit microseconds so clusters scale ~linearly.

@LLMJunky @alexocheema @davidweiss A M3 Ultra with 512gigs costs >25k. Good luck finding 6 even. Beside what you wanna do with a 3tb model and 80gb cable connection, generate 1t/s?

@stev_builds @alexocheema @davidweiss the counter argument to this would be where else are you going to find 3TB of "vram?" for under $75K?

@DeItaone RIP for them who relies on disc's as they have poor internet connection. In my summer house I have 20mb internet and even updates took ages

PLAYSTATION: TO END PHYSICAL DISC PRODUCTION FOR NEW GAMES STARTING JANUARY 2028

@0x99Gohan @kylobtc Waiting for an airdrop

@kylobtc what? Over 50% of supply staked? That's loyal

Injective has around 100M $INJ circulating 56.8M of that is already staked That's more than half the supply helping secure the network instead of just sitting liquid Market cap is only around $521M here For a chain trying to power onchain finance, tokenized assets, orderbooks and institutional-grade markets, that setup is pretty interesting High staking, low float pressure, and a network built around real financial activity

@alexocheema @davidweiss Why the hack you would want cluster 6 M3 Ultras? You can combine 5 devices together, enjoy multiple bottlenecks. If you want to spend that kinda money there are way better solution for inference

@davidweiss It's limited by numbers of Thunderbolt 5 ports. You can cluster 4 x MacBook Pro or 4 x M4 Pro Mac Mini or 6 x M3 Ultra Mac Studio.